CONCEPTS

What “evidence-grounded” means in Candor

Most AI persona tools generate a person from a prompt. Candor doesn’t. Every audience, every persona, and every claim in a report traces back to specific evidence. Here’s what that means in practice and how to read the signals.

The principle

A confident-sounding AI persona is dangerous if you can’t tell whether their answer is grounded in real-world evidence or generated from thin air. Confidence without provenance is indistinguishable from confabulation. Candor’s answer is to never present a fact without a tag describing where it came from, and to never claim more confidence than the evidence supports.

This shows up everywhere: every signal in an audience, every field in a persona, every claim in a synthesis report carries a provenance tag.

The five tags

Grounded

The strongest tag. The claim cites exactly one specific source — a passage from your uploaded research, a web page, or another concrete artifact. You can click through to see the original. If a persona’s decision rule is tagged grounded, there’s a real document or page that says so.

Inferred

The claim synthesises across two or more sources. No single citation captures it; the conclusion emerges from the pattern. Still well-supported, just not pinned to one quote. Most belief-memory entries are inferred — beliefs rarely come from a single sentence.

Calibrated

Derived from validated behavioral or population distributions, not from a specific source you can cite. Used when the question is something like “what range of openness is plausible for this audience?” — the answer comes from peer-reviewed personality research, not from a particular interview transcript.

Sampled

Drawn from a calibrated distribution. The persona’s specific OCEAN values and bias intensities are sampled — each one is a specific value pulled from the calibrated range that defines their archetype. Not arbitrary; the distribution it came from is grounded in research, but the specific value is one realisation among many possible.

Weak confidence

An explicit hypothesis with no source backing yet. Marked this way deliberately, not by accident. Candor uses this when a claim is needed to make the persona coherent but the evidence to support it isn’t available. Treat weak-confidence claims as things to test, not things to trust.

Why five tags and not just three

A simpler scheme would be has source / no source / made up. The reason Candor uses five is that the difference between calibrated (grounded in scientific distributions) and sampled (a specific value drawn from those distributions) is real and useful. Both are research-backed; neither is a citation. Collapsing them would lose information.

The tags are also colour-coded in the UI so you can scan a persona profile and see at a glance how much of it is citation-grade vs. distribution-derived vs. hypothesis.

Where you’ll see grounding in the product

  • Audience signals — each signal carries a provenance dot (grounded, inferred, calibrated, or hypothesis) plus a source pill linking to a document or web page.
  • Persona profile — every memory field and B2C/B2B attribute carries a provenance badge. The audience-quality % at the top rolls all of these into a single number.
  • Synthesis reports — the evidence log section links every theme, tension, and claim back to the source quotes that supported it.

How to use grounding when reading

Three habits that pay off:

  • Don’t treat all claims as equal. A grounded decision rule and a weak-confidence one belong in different categories of confidence in your head.
  • Click through on grounded signals. They’re fast to verify, and seeing the source builds your intuition for what the audience actually looks like.
  • Test weak-confidence claims explicitly. When a persona’s objection is tagged weak confidence and the synthesis depends on it, that’s a candidate for a follow-up live interview to confirm or contest.

What this doesn’t solve

Grounding tells you where a claim came from. It doesn’t tell you whether the source was correct. Web evidence can be wrong. Uploaded research can be incomplete. A grounded persona built from biased sources will reflect that bias. Provenance is a tool for healthy scepticism, not a replacement for it.

Where to go next

Common questions

Look at the provenance tag. Grounded means there's a specific source you can click through to. Inferred means it's synthesised across multiple sources. Calibrated means it came from peer-reviewed research distributions. Sampled means it's a specific value drawn from one of those distributions. Weak confidence means it's an explicit hypothesis with no source backing yet, and Candor marks it that way deliberately so you know. A grounded decision rule and a weak-confidence one shouldn't carry the same weight in your head.

Provenance tracks where a claim came from, not whether the source was correct. Web evidence can be wrong. Uploaded research can be incomplete or biased. A grounded persona built from biased sources will reflect that bias. That's why provenance is a tool for healthy scepticism, not a replacement for it. The fix is to click through on grounded signals you find surprising and verify the source. Strong evidence sources produce strong personas. Weak sources produce weak ones, no matter how confident the model sounds.

Because the difference between calibrated (grounded in scientific distributions) and sampled (a specific value drawn from those distributions) is genuinely useful. Both are research-backed, but they answer different questions. Calibrated tells you the plausible range. Sampled tells you which value got picked. Collapsing them would lose that information. The same logic applies to splitting weak confidence (explicit hypothesis, no source) from inferred (well-supported across multiple sources). Five tags map the actual gradient of how confident you should be.

Yes. Your uploaded research gets higher priority than web sources during signal extraction. The reasoning: you uploaded those documents because they're directly relevant to your audience, while web search casts a wide net and pulls in adjacent or partial matches. The audience-quality score on each persona reflects this weighting. The more of your audience is built from documents you trust, the higher the quality score, and the more confident you can be in the persona's claims.

Three places. On every audience signal (a coloured dot plus a source pill linking to the document or web page). On every persona memory field and B2C/B2B attribute (a provenance badge next to the value). In synthesis reports, the evidence log section links every theme, tension, and claim back to the source quotes that supported it. Scanning provenance is the fastest way to read a persona profile or a report critically. The colour-coded tags let you see at a glance how much is citation-grade versus hypothesis.

More FAQs →

Candor is in development.

Be the first to know when it launches.

No spam. Just a note when Candor is ready. Powered by Highline Beta.